Thesis Defense - Aurora Rossi (Université Côte d'Azur, Centre Inria) - 25/09/2025

Soutenance de Thèse - Thesis Defense Computational Methods and Analysis of Temporal Networks - Méthodes Computationnelles et Analyse des Réseaux Temporels Aurora Rossi (Université Côte d'Azur, Inria) This thesis, supervised by David COUDERT, focuses on the application of graph theory, in particular temporal graphs (i.e., graphs that change over time), to other sciences such as neuroscience. She is part of the COATI team, a joint team between the Inria Research Center at Université Côte d'Azur and the i3S laboratory. She also collaborates with Inria's CRONOS team. In the summer of 2023, Aurora ROSSI participated in the Google Summer of Code program and JuliaCon 2023 at the Massachusetts Institute of Technology. In 2025, she spent two months at the Institute of Science Tokyo in Japan with the RISE Academy's DocWalker international mobility program. Cette thèse, suprervisée par David COUDERT, porte sur l'application de la théorie des graphes, en particulier des graphes temporels (c'est-à-dire des graphes qui changent au fil du temps) à d'autres sciences telles que les neurosciences. Elle fait partie de l'équipe COATI, une équipe commune entre le Centre de Recherche Inria d'Université Côte d'Azur et le laboratoire i3S. Elle collabore également avec l'équipe CRONOS d'Inria. A l'été 2023, Aurora ROSSI a participé au programme Google Summer of Code et à la JuliaCon 2023 au Massachusetts Institute of Technology. En 2025, elle a passé 2 mois à l'Institute of Science Tokyo au Japon avec le programme de mobilité internationale DocWalker de l'Académie RISE. Abstract: This thesis develops computational methods for the analysis of temporal networks, with an emphasis on applications to neuroscience. Temporal graphs provide a natural representation for dynamic systems in which interactions evolve over time, such as functional connectivity in the human brain. As a relatively recent area of study, temporal graph theory still requires the development of dedicated tools and generalizations of classical static methods. To address this gap, new metrics are proposed to extend classical notions, such as clustering coefficient, path length, and small-worldness, into the temporal domain. These are designed to capture both local and global structure while respecting the causality imposed by time. A novel null model, the Random Temporal Hyperbolic model, is introduced to simulate the small-world property observed in brain dynamics while randomizing other topological features. This provides a meaningful baseline for the statistical evaluation of temporal connectivity patterns. A machine learning framework is also introduced to identify task-relevant subnetworks from temporal connectivity data. This approach integrates Shapley values to quantify the contribution of each subnetwork to the model’s predictions. Applied to fMRI recordings during naturalistic stimuli, the framework reveals interpretable patterns in how different brain regions support narrative comprehension, aligning with current cognitive neuroscience theories. To support and generalize these computational approaches, I contributed to developed GraphNeuralNetworks.jl, an open-source Julia package for building and training graph neural networks on static, temporal, and heterogeneous graphs. This flexible and performant software enables rapid experimentation with modern graph-based learning models and is used here in the context of both brain decoding and traffic forecasting. Altogether, this thesis introduces a collection of reusable mathematical, algorithmic, and software tools designed for modeling and analyzing temporal graphs, while contributing to our understanding of the brain’s dynamic architecture. Though they were initially developed in neuroscience, these methods can be applied to a broad range of disciplines that deal with evolving relational data. 👉 En savoir + sur les thèses financées par l'EUR Systèmes Numériques pour l'Humain : https://ds4h.univ-cotedazur.fr/recher...